---
license: mit
language:
- en
library_name: transformers
pipeline_tag: feature-extraction
base_model: BAAI/bge-small-en-v1.5
tags:
- information-retrieval
- mteb
- scientific-retrieval
---

# BGE Small Structural Separator

This is the complete Transformers checkpoint for the H216 Shared Structural
Separator Field Encoder. It is compatible with `AutoModel` and `AutoTokenizer`.

H216 starts from the immutable
`BAAI/bge-small-en-v1.5@5c38ec7c405ec4b44b94cc5a9bb96e735b38267a`
checkpoint and learns exactly one 384-value input-embedding row: token
`[unused2]`, vocabulary ID 3. All other model parameters are unchanged.

Document formatting inserts `[unused2]` before a nonempty title and before each
punctuation-delimited sentence. The document is encoded once and its normalized
CLS vector is stored. Queries use ordinary tokenizer formatting with no
instruction. Retrieval is exact cosine over one query vector and one document
vector; there is no fusion, routing, reranking, expansion, or protected
candidate frontier.

```python
from transformers import AutoModel, AutoTokenizer

model_id = "thu-nmrc/bge-small-structural-separator"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id)
```

The repository includes `separator_row.pt`, the deterministic training manifest,
and `export_manifest.json` with SHA-256 hashes for the complete checkpoint.
Training used 4,096 S2ORC citation-context/abstract pairs and no BEIR labels,
validation selection, benchmark negative mining, or post-training tuning.

The bounded research claim is a learned document-side structural separator
adaptation within the dense bi-encoder family. It is not claimed as a replacement
for all dense retrievers. Full evaluation details, training code, overlap audits,
and disclosed regressions are maintained at
https://github.com/thu-nmrc/bge-small-structural-separator.

The frozen BEIR8 gate improves macro NDCG@10 by +0.012705 over the matched plain
BGE control. A later preregistered six-task `MTEB(eng, v2)` Retrieval extension
does not support broad transfer: macro NDCG@10 changes by -0.011038 and H216 wins
two of six tasks. Across all ten official Retrieval tasks the macro delta is
+0.002855, but claims remain bounded to the BEIR8 structure-native evidence.